close

Вход

Забыли?

вход по аккаунту

?

SAR-based estimates of the size distribution of lakes in Brazil and Canadaa tool for investigating carbon in lakes.

код для вставкиСкачать
AQUATIC CONSERVATION: MARINE AND FRESHWATER ECOSYSTEMS
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
Published online in Wiley InterScience
(www.interscience.wiley.com) DOI: 10.1002/aqc.854
SAR-based estimates of the size distribution of lakes in Brazil
and Canada: a tool for investigating carbon in lakes
KEVIN H. TELMERa,b,* and MAYCIRA P.F. COSTAc,d
a
School of Earth and Ocean Sciences, University of Victoria, Victoria, BC, Canada
b
University of Campinas, Campinas, SP, Brazil
c
Department of Geography, University of Victoria, Victoria, BC, Canada
d
National Institute of Space Research (INPE), São Jose´ dos Campos, SP, Brazil
ABSTRACT
1. The size of lakes and the size distribution of lakes are important parameters controlling lake
function, and how lakes interact with landscapes, the atmosphere and ecosystems. A baseline digital
database of lakes could be used to improve understanding of lake function, to extrapolate lake
information to regional and global scales, and as a basis for detecting future changes to lakes.
2. This paper examines the capabilities of synthetic aperture radar (SAR) imagery produced by the
Japanese Earth Resources Satellite (JERS-1) to map the number and size distribution of lakes in
western Arctic Canada, Central Canada and the Pantanal (Brazil).
3. For the Arctic and Pantanal, the total area found within one lake size category increases
towards smaller lakes. The opposite was true for the area in Central Canada. The number of lakes in
the smallest size category, 0.01 to 0.1 km2, was underestimated for all regions owing to the resolution
of the mosaics } 100 100 m. The number of large lakes in the Pantanal was over-estimated
through confusion with intermittent floodways that are scrubby grasslands and bare sand in the dry
season and which exhibit low backscattering and therefore appear dark like lakes.
4. The lake distributions were combined with existing data to produce preliminary regional
estimates of carbon accumulation. Lakes may accumulate as much as 1.7 and 1.3 t C km2 yr1 for
the Arctic Canadian and Central Canadian areas, respectively. No estimates were produced for the
Pantanal because there are no applicable data on carbon accumulation rates available for that
region.
Copyright # 2007 John Wiley & Sons, Ltd.
KEY WORDS: Lakes; Canada; Brazil; SAR; radar; JERS-1; ALOS; carbon accumulation; size distribution;
Pantanal; Boreal
*Correspondence to: K.H. Telmer, School of Earth and Ocean Sciences, University of Victoria, Victoria, BC, Canada.
E-mail: ktelmer@uvic.ca
Copyright # 2007 John Wiley & Sons, Ltd.
290
K.H. TELMER AND M.P.F. COSTA
INTRODUCTION
Carbon accumulated in lake sediments has been recognized as an important component of the global
carbon cycle and a significant sink for atmospheric CO2, but has been remarkably poorly studied compared
with other carbon sinks of similar magnitude such as carbon storage in soils, peat, terrestrial biomass or the
oceans. Mulholland and Elwood (1982) were perhaps the first to recognize the importance of carbon
storage in lakes with their estimate that globally, lakes accumulate 0.02 Pg C yr1. Others have since
modified this estimate. For example, Dean and Gorham (1998), largely using lakes in Minnesota and the
eastern United States, estimated that lakes globally accumulate as much as 42 Tg of organic carbon per
year. Einsele et al. (2001) claim that 70 Tg of organic and inorganic carbon are buried each year in lakes,
representing 25% of the atmospheric carbon buried in the oceans annually, which has been estimated to be
250 Tg by Schlünz and Schneider (2000). Pajunen (2000) and Kortelainen et al. (2004) report the findings of
a detailed survey of sedimentary carbon in 122 lakes in Finland. They conclude that within Finland, aquatic
ecosystems, despite covering only 10% of the land area, contain the second largest carbon stocks
(19 kg C m2) after peatlands (72 kg C m2), and significantly exceed stocks in the forest soil (uppermost
75 cm; 7.2 kg C m2) and woody biomass (3.4 kg C m2). The Finnish estimate extrapolated over the boreal
region gives a total carbon pool in lakes 19–27 Pg C, large, but significantly lower than the previous modelbased estimates.
These studies represent important progress but understanding and quantifying the way that carbon cycles
and accumulates in lakes remains relatively poorly constrained. Particularly lacking are: (i) estimates of the
variation of carbon accumulation in lakes from different geomorphologic terrains and climates (particularly
tropical), and (ii) good estimates of the number and size distribution of lakes for these different
geomorphologic terrains and climates. This latter point is particularly important because carbon dynamics
in lakes vary with lake size in a variety of ways.
For example, Fee and Hecky (1992) designed an experiment called the ‘Northwest Ontario Lake Size
Series’ (NOLSS) to investigate the role of lake size on various processes in lakes. They selected six lakes that
ranged in size by six orders of magnitude but were otherwise believed to be very similar in their setting in
terms of vegetation, climate, and geology. Using this design, Fee et al. (1992) illustrated that lake size affects
phytoplankton photosynthesis. Also using the NOLSS lakes, Kelly et al. (2001) investigated the role of lake
size on epilimnetic carbon dioxide production. They found that lake size is an important factor with small
lakes emitting more CO2. Jonsson et al. (2003) investigated CO2 in clearwater and humic lakes in Sweden
and found strong correlations between the dissolved organic carbon (DOC) content in lakes and CO2
supersaturation, and also that sediment respiration is most significant in shallower lakes. Pajunen (2000)
and Kortelainen et al. (2004) have shown that smaller lakes typically accumulate carbon the fastest.
Moreover, the role that lakes play in landscape carbon cycling has recently been recognized to be more
important that previously thought. Pace et al. (2004) have demonstrated that terrestrial carbon (exogenous
carbon) substantially subsidizes food webs in lakes, while Algesten et al. (2004) investigated 79 536 lakes in
Sweden and found that mineralization of terrestrially derived organic carbon in lakes is an important
regulator of organic carbon export to the sea and may affect the net exchange of CO2 between the
atmosphere and the boreal landscape. They highlighted the role of hydrology, of which lake size and
distribution is an important factor in controlling lake carbon dynamics.
In order to improve understanding of how carbon cycles and accumulates in lakes, and to increase the
abilities to transfer field-based knowledge to large areas, we plan to use the all-weather and vegetation
penetrating capabilities of L-band synthetic aperture radar (SAR) to produce a digital database of the
world’s lakes: a census of lakes from a singular time period that quantifies lake size and spatial distribution
for the globe that is readily queried and accessible.
Because it uses L-band microwave energy, has high spatial resolution, is multipolarimetric, and operates
in all weathers and by night and day, the ALOS-PALSAR instrument, recently launched by the Japanese
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
SAR-BASED ESTIMATES OF THE SIZE DISTRIBUTION OF LAKES
291
Aerospace Exploration Agency (JAXA), is particularly well adapted for this task. L-band can see
through clouds and sparse vegetation (Leckie and Ranson, 1998) and is less sensitive to water surface
roughness than shorter wavelengths such as the C-band used by Canada’s RADARSAT (Martin,
2004). These qualities will allow ALOS to differentiate water from land under the greatest diversity of
conditions. Because L-band penetrates vegetation, ALOS should easily identify small lakes covered or
fringed with short aquatic vegetation (Costa et al., 2002) such as those in the Brazilian Pantanal discussed
by Costa and Telmer (2007). Its multipolarimetric abilities may improve target recognition
in cases where vegetation penetration is needed. The high spatial resolution of the global mosaics
planned to be produced by ALOS should allow even small lakes to be imaged. Imaging small lakes is
critical because of their importance in carbon cycling and accumulation, and because they are also the
ones most susceptible to climate change (Schindler et al., 1996; Smith et al., 2005). An important
capability of radar satellite imagery for this project is its all-weather and night/day capabilities. ALOS,
for example, will be able to collect relatively time-contiguous data for large regions. This is critical for
capturing all of a region’s lakes under the same climatic conditions. This will greatly reduce the errors
in mapping small lakes whose characteristics change strongly with the seasons. Emergent aquatic
vegetation cover, for example, can wax and wane from the dry season to the wet, and the size of
shallow fluvial lakes can vary significantly through the seasons as well as interannually. Mapping lakes for a
large region using a patchwork of imagery from different seasons and years reduces the quality of the
database.
Several other lake censuses have been completed in the last 50 years, with the most recent and
comprehensive completed by Meybeck (1995). Using rainfall/runoff relationships for the globe,
Meybeck extrapolated existing lake surveys and registers to a global scale. This was an important
accomplishment, and an excellent starting place for our proposed work. However, Meybeck’s survey, and
those before it upon which his was based, do not allow some important types of analysis to be performed.
For example, these earlier surveys do not come from a single data source produced by a single method
nor from a single time period. A database that did meet these criteria would eliminate inter-study biases
that result from merging data sets that were collected over many years by different researchers for a variety
of purposes. These older surveys are also static data sets (essentially tables) that do not allow much
querying or analysis. The digital database that we propose would allow lakes to be easily classified in many
ways (any size classes for example) and would allow queries to be performed at any spatial scale}locally
to globally. This would help to develop a better understanding of what controls lake distribution and
other lake processes for different regions and at different scales, and in turn would allow carbon fluxes to be
calibrated and queried by region and globally. Finally, perhaps most importantly, a database of lakes
produced from a single season would provide a snapshot of lake distribution frozen in time that can form
the basis upon which to detect changes to the world’s lakes in the future}for example, changes induced
by shifts in climate or by human development. The PALSAR sensor on ALOS makes such a comprehensive
survey possible.
The remainder of the paper presents some promising results for some prototype areas based on existing
JERS-1 mosaics. These results do provide estimates of lake size distribution for the prototype areas but
they are also useful for illustrating methodological problems which the ALOS acquisitions will need to
overcome. In addition, these results begin to provide a radar based database from the 1990s that can be
compared with future ALOS data for the purpose of detecting change.
STUDY AREA
This paper discusses results for two 200 200 km areas in Canada and one 200 200 km area in the
Pantanal (Figure 1). These areas were chosen because they contain a wide variety of lake sizes and types
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
292
K.H. TELMER AND M.P.F. COSTA
Figure 1. Synthetic aperture radar (SAR) images of the study areas, represented by black dots, generated from the Japanese Earth
Resources Satellite (JERS-1)–lakes are dark. The upper left image was acquired in 1997 and lies on the border of the Northwest and
Nunavut Territories in the Canadian Arctic. The largest lake in the image is called Napaktulik Lake which means ‘the place where
there are trees’ (formerly it was called Takijug Lake); its north shore touches the Arctic Circle. The upper right image was acquired in
1997 and lies on the border of the provinces of Manitoba and Ontario in Central Canada. The large ‘messy’ lake in the lower right of
the image is called Island Lake. The lower left image was acquired in 1993 and is of the Nhêcholandia region of the Brazilian Pantanal.
It lies just to the east of the Brazilian–Bolivian border. The boxes on the world map represent the full extent of the prototype areas to
be investigated for the global lakes project under the ALOS Kyoto Initiative. There are no historical JERS-1 mosaics for the Zambezi
Basin region and so it is not included in this paper. Each image is 200 200 km (40 000 km2), and is composed of 2000 2000 pixels,
each pixel has dimensions of 100 100 m. The boxes on the images are the areas shown in Figures 2, 3 and 4. This figure is available in
colour online at www.interscience.wiley.com/journal/aqc
whose spatial distribution varies depending on ecosystem characteristics (Abdon et al., 1998; Costa
and Telmer, 2006). They represent temperate and tropical climates as well as areas that are relatively
well studied (Boreal) and poorly studied (Pantanal). These characteristics are useful for testing the
robustness of the census to various conditions and finding weaknesses in the methodology. In addition,
the chosen areas are representative of, and overlap with, relevant field-based studies on the biogeochemistry
of landscapes and lakes such as the BOREAS project (Hall, 1999), the NOLSS project (Fee
and Hecky, 1992), and work in the Pantanal (Costa and Telmer, 2006). Figure 1 also shows the
larger areas that form the prototype areas for the initial ALOS data stream–Canada, The Pantanal and
southern Africa.
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
SAR-BASED ESTIMATES OF THE SIZE DISTRIBUTION OF LAKES
293
DATA
The data comprise SAR imagery produced from the Japanese Earth Resources Satellite JERS-1 by the
National Space Development Agency of Japan (NASDA), now called the Japanese Aerospace Exploration
Agency (JAXA). JERS-1 was a polar orbiting satellite that produced L-band imaging radar. It was in
operation between April 1992 and October 1998. Satellite characteristics: polar orbit (97.678); orbital
altitude: 568 km; repeat cycle: 44 days; swath width: 75 km. SAR characteristics: L-band (23.5 cm/
1.275 GHz); polarization: HH; off-nadir angle: 35 8; spatial resolution: 18 18 m.
The data used here were acquired during the summer (July to August) 1998 for Canada, and the dry
season (July to September) 1993 for Brazil. The raw data (18 m resolution) was processed by the Global
Forest Mapping Project to produce mosaics with 100 m resolution (Rosenqvist et al., 2000). The images are
grey scale in which the brightness of the image represents the amount of microwave energy (radar)
backscattered (returned) to the satellite. Each pixel in the image can have a digital number of 0 to 255 (8-bit
data).
The 100 m mosaics were used here to estimate the abilities of the 50 m mosaics proposed to be produced
from ALOS-PALSAR and also to investigate the usefulness of the existing JERS-1 data to be used for
comparison with future ALOS data. Such a comparison could provide decadal-scale change analysis.
The grey tones in the digital mosaics represent the strength of backscattering from different ground cover
in the terrain because different objects on the earth’s surface scatter radar differently (Ulaby et al., 1982).
For example, flooded forests such as those in the Amazon floodplain during the high-water season strongly
reflect radar (double-bounce mechanism) (Hess et al., 1995; Costa, 2004) and so appear as very light tones
in the images. Tree canopies reflect a medium amount of radar (volume scattering) (Dobson et al., 1996;
Costa, 2004) and so appear as mid-tones; and most importantly for this paper, water typically reflects very
little radar owing to specular reflection and so lakes appear as dark tones in the images (Lewis, 1998).
However, the amount of backscattering from water surfaces such as lakes is strongly influenced by
environmental characteristics such as the roughness of the water surface, presence of emergent vegetation,
ice conditions, surrounding relief, and surrounding ground cover (Lewis, 1998). Waves caused by wind or
other processes cause incident radar pulses to be backscattered by a mechanism called ‘Bragg scattering’.
The amount of backscattering is related to the incidence angle of the radar, the wavelength of the radar,
and the wavelength of water waves (Martin, 2004). Under similar conditions, backscattering will be lower
for larger radar wavelengths. This means that L-band radar will backscatter less than C-band radar for a
water body with the same surface roughness. This in turn means that L-band radar is less sensitive to
weather conditions than C-band for observing water, making it a better choice for differentiating water
from land.
With regards to the interaction of radar and emergent aquatic vegetation that may surround lakes,
scattering is expected to be caused by the elements of the vegetation canopy, a mechanism called ‘volume
scattering’ (Ulaby et al., 1982), which results in higher backscattering compared with that from flat water
surfaces. With a wavelength of 30 cm and an incidence angle of 358 the L-band JERS-1 data are expected to
penetrate deeply short emergent aquatic vegetation, because of the relationship between size of wavelength
and the size of canopy structures (Costa, 2004). This indeed was the case for Pantanal lakes observed by
Costa and Telmer (2006). They found that the low backscattering values for JERS-1 (14.5 dB) suggested a
mix of volume scattering (canopy) and specular scattering (off the water) as the most probable scattering
mechanisms. However, it must be noted that Hess et al. (1995), Novo et al. (2002) and Costa (2004) have
observed a strong double-bounce mechanism for L-band radar with tall (2 m) dense aquatic vegetation in
the Amazon, so the transparency of emergent aquatic vegetation to L-band radar may not be simply
assumed. Perhaps multi-polarimetric data such as will be produced by ALOS will assist in better
understanding this. The usefulness of multipolarimetric imagery for wetlands mapping was already
demonstrated by Hess et al. (1995) (shuttle C and L multi-polarization over Amazon), Simard et al. (2002)
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
294
K.H. TELMER AND M.P.F. COSTA
(L-band multi-polarization over mangroves in Africa), and Costa (2004) (JERS-1 and RADARSAT over
Amazon–not multi-polarimetric only multi-band).
METHOD
A histogram of each image was created. Each image was then enhanced (linearly) to stretch the digital
numbers to the full range possible (0–255). This aided in the visualization of the data and ultimately in
selecting a threshold. A sample of the original images, the enhanced images and the histograms of both the
original and enhanced images is shown in Figures 2, 3 and 4. The values of pixels that represent lakes and
those that represent land were queried, by selecting areas that were known to be only land or only lake in
the enhanced images. For example, to ascertain the digital numbers (DN) that represented water,
approximately 100 pixels from the centre of a large lake were selected, with a similar procedure for land.
Thresholds that would effectively separate land from water were determined and the images were then
classified into two classes, lakes and land, using a single threshold to separate them. The thresholding
technique is a simple and computationally efficient method for defining boundary limits of lakes using SAR
imagery compared with classification techniques in which definition of training sites is required (Kozlenko
and Jeffries, 2000).
Figure 2. (a) A 24 24 km unenhanced zoom from the JERS-1 SAR image of the Northwest Territories as shown by the box in
Figure 1. (b) The same but linearly enhanced. (c) The histograms of the digital numbers (0 to 255) from the full 200 200 km area;
shaded is unenhanced; bars are for the linear enhancement; y-axis is frequency from 0 to 3%; and the diagonal line illustrates how
the original values were stretched.
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
SAR-BASED ESTIMATES OF THE SIZE DISTRIBUTION OF LAKES
295
Figure 3. (a) A 24 24 km unenhanced zoom from the JERS-1 image of the Pantanal as shown by the box in Figure 1. (b) The
same but linearly enhanced. (c) The histograms of the digital numbers (0 to 255) from the full 200 200 km area; shaded is unenhanced;
bars are for the linear enhancement; y-axis is frequency from 0 to 3%; and the diagonal line illustrates how the original
values were stretched.
In terms of choosing thresholds, several approaches were tried to find the most effective one. First,
with, the aim of being able to use one threshold for all three study areas, the data were linearly enhanced
to normalize the lowest values to zero. Second, assuming that the bimodal distribution observed in the
two Canadian study areas represented water and land peaks, a unique threshold for each area that
was equidistant between bimodal peaks was applied. For the Pantanal, in this case, there are weak high
and low peaks at the end and beginning of the range of the data and so a DN equidistant between these
was selected. Third, a tuning approach was adopted where unique DN thresholds were applied and
the results were evaluated, and then the threshold was adjusted until a satisfactory result was
obtained. Evaluation was completed by overlaying images classified into lakes and land on the enhanced
imagery to check visually whether or not lakes were being correctly distinguished. In the end, this
visual procedure does produce satisfactory results; however, it may be possible in future to evaluate
classifications more quantitatively by comparing results with other well characterized data sets and ground
truth.
Applying a single threshold to all three areas did not produce satisfactory results. It led to one area
being properly classified and the other two not. The equidistant approach was also unsatisfactory as it
generally underestimated the areas and numbers of lakes, so at this stage a custom fit for each area
was performed. Figure 5 shows the overlay of areas classified as lakes onto enhanced imagery for the area
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
296
K.H. TELMER AND M.P.F. COSTA
Figure 4. (a) A 24 24 km unenhanced zoom from the JERS-1 image of the Manitoba-Ontario border as shown by the box in
Figure 1. (b) The same but linearly enhanced. (c) The histograms of the digital numbers (0 to 255) from the full 200 200 km area;
shaded is unenhanced; bars are for the linear enhancement; y-axis is frequency from 0 to 3%; and the diagonal line illustrates how
the original values were stretched.
of the boreal forest that lies on the border of the Canadian provinces of Manitoba and Ontario (Central
Canada). It shows the result of using the latter two approaches. The equidistant approach resulted in a low
threshold and missed many small lakes and underestimated lake size by overestimating shoreline thickness.
The higher customized threshold produced more satisfactory results–few obvious lakes were missed and
shorelines were not exaggerated. The final DN thresholds applied to the enhanced images were Arctic
area ¼ 80; Boreal area ¼ 130; Pantanal ¼ 100: In the end, linear enhancement of the images was needed
only for optimal visualization.
Once satisfied with the results, shorelines were automatically drawn around lake areas to form
closed polygons using an edge-detection algorithm in the software package Geomatica by PCI Geomatics
Inc. The number of lakes and the area of each was then determined and tabulated by analysing
the polygons. Figure 6 shows the shorelines that were created for a part of the Arctic study area that lies
on the border of the Northwest and Nunavut Territories. The lakes were then grouped into the following
size categories: 0.01, 0.1, 1, 10, 100, 1000 and 10 000 km2. This produces the size distribution of the lakes
for the area covered for each image. Finally, an estimate of how much carbon might accumulate in these
lakes was made based on accumulation rates published in the literature and on the observed size
distribution of the lakes.
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
SAR-BASED ESTIMATES OF THE SIZE DISTRIBUTION OF LAKES
297
Figure 5. Overlay of areas classified as lakes onto enhanced imagery. Zooms of the centre of the main image show the results
of using two different digital number thresholds. A low threshold was used for the upper panel but this missed many small
lakes and underestimated lake size by overestimating shoreline thickness}this is most obvious as the dark boundaries around islands.
A higher threshold was used for the lower panel and this produced more satisfactory results}few lakes were missed and
shorelines were not exaggerated. These issues will be less of a problem with the higher resolution ALOS imagery due to begin to be
produced in 2006. This figure is available in colour online at www.interscience.wiley.com/journal/aqc
RESULTS AND DISCUSSION
As shown by the histograms in Figures 2, 3 and 4, the range and distribution of backscattering for the three
regions are different because of the different composition of the ground surfaces. Statistics for the raw and
enhanced images are listed in Table 1. The two Canadian study areas have smooth bimodal distributions
whereas the Pantanal has more uneven and unimodal distribution. Of the two Canadian study areas, the
Boreal area in Central Canada has higher contrast with a wider range of raw values and greater separation
between the two groups of data (the two peaks on the histogram). In terms of radar backscattering, densetall vegetation is typically bright, bare soil and sand have intermediate backscattering, and water is dark
(Ulaby et al., 1982), and so these results make sense from what is known about the study areas. The Boreal
area is the most heavily forested and so has the highest median and mean backscattering values. The Arctic
study is less vegetated (tundra) and has scrubby and low but abundant vegetation, so it has lower median
and mean DN than the Arctic study area. Both Canadian study areas have a low DN water peak that
represents water with relatively few areas with intermediate values. This suggests that the two Canadian
areas have little bare soil or sand. The Pantanal, on the other hand, has a similar overall range to that of the
Boreal area suggesting both dense vegetation and water, yet also has many intermediate values. This also
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
298
K.H. TELMER AND M.P.F. COSTA
Figure 6. In red are the shoreline polygons created for the lakes of this 200 200 km JERS-1 scene of an area of Canada’s NWT. All
lakes larger than 10 km2 are highlighted in white. Two full-resolution zooms (24 24 km) of the centre of the image with the shorelines
and without are shown to the right.
Table 1. Statistics on digital numbers (DN) for the raw and enhanced images of the three study areas
Number of pixels
Median DN raw
Median DN enhanced
Mean DN raw
Mean DN enhanced
StDev raw
StDev enhanced
Shape
Western Arctic
Central Canada
Pantanal
4 000 000
84
136
81.6
128.2
22.6
64.9
bimodal
4 000 000
110
170
102.8
151.2
27.4
69.7
bimodal
4 000 000
82
112
83.3
114.4
28.8
68.6
unimodal
makes sense as the Pantanal contains many lakes, corridors of forest, and savannah-like grasslands and
sandy plains called ‘vazantes’.
Although the imagery generally follows expectations for what is known about the targets, results are not
perfectly consistent from region to region. For example, some of the water in the Pantanal is ‘darker’ (lower
DNs) than in Canada. This could be due to differences in original image quality caused by ambient
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
SAR-BASED ESTIMATES OF THE SIZE DISTRIBUTION OF LAKES
299
conditions, such as wind, or perhaps due to the radiometric calibration intrinsic to radar images (see
Chapman et al., 2002). Whatever the reason, this meant that a single threshold could not be applied to
separate water from land for the different areas. Similar problems using SAR for determining lake area
were reported by different researchers, and summarized in Lewis (1998), as a result of wind-roughened
water surfaces, emergent aquatic vegetation, and ice conditions (for cold regions)–all result in higher radar
return.
The upper and lower panel of Figures 7 and 8 illustrates the estimates of the size distribution and carbon
accumulation rates of the lakes determined for each study area. Carbon accumulation rates were calculated
using the observations made by Kortelainen et al. (2004). High and low estimates are shown corresponding
to the two approaches used for choosing thresholds discussed above. The data are listed in Table 2. The
lake size ranges are given both in square kilometres and in terms of pixels to aid in evaluating resolution
issues. Each pixel is 100 100 m or 0.01 km2. Small lakes that are represented by single or a low number of
pixels are difficult to resolve because a single pixel may be capturing a mix of both land and water, and will
appear brighter than a pure water pixel. This is an unavoidable result of several factors, including the
averaging method applied to create the mosaics at 100 100 m resolution (Chapman et al., 2002).
Therefore, the numbers of lakes identified in the smallest size class (0.01–0.1 km2 or 1–10 pixels) are
uncertain and are probably low.
Thresholds based on expectations for pure water would not capture these lakes. This is why the threshold
ultimately chosen for the separation of water from land fell on the rising limb of the ‘land’ peak in the
histograms for Canadian study areas. Visually, it was obvious that lower thresholds missed many small
lakes but it appears that even the higher thresholds ultimately chosen are also not capturing the true
number of lakes in the smallest size category. The under-representation of the smallest lakes is obvious
from the shape of the power laws.
The shape of the power law that describes the size distribution of the lakes, the lower panel of Figure 7,
the log–log plot, shows power laws with relatively straight lines until the smallest size class is reached–there
the slope of the lines decreases. In Meybeck’s (1995) lake survey and in the many studies that he used in his
summary, the power law relationship observed for lakes held across six orders of magnitude and it was
speculated that it holds for smaller lake size classes as well. Here, if it is assumed that the decrease in slope
in the lines shown in Figure 7 is due to imagery resolution and that the power law relationship continues on
to the smaller size class, then the number of lakes in the smallest size class can be estimated simply by
extending the power law from the other size categories. This was the approach used for tundra ponds (see
discussion in Meybeck (1995)). The equations for the relationships are given in the caption of Table 1. The
increases in the estimate of the number of lakes for the 0.01–0.1 km2 size class caused by extrapolating
the power law are large (numbers in Table 1 in italics). For example for the Arctic region of Canada, the
number of lakes determined for the 0.01–0.1 size class is 40 000 for a 100 000 km2 area. Extrapolating the
power law derived from the larger size categories to the small category increases this number to 200 000
lakes–five times more.
Whether or not this is reasonable can only be determined by fieldwork for selected areas and/or by using
higher resolution imagery. According to the power law relationships, the 100 m pixels from the JERS-1
mosaics do appear effective in identifying lakes of the size class 0.1–1 km2–those containing 10 pixels or
more. This agrees with the findings of Lewis (1998), that the diameter of the lake should be at least three
times the spatial resolution of the SAR system for lakes to be detected (i.e. an area of nine pixels).
The proposed ALOS mosaics for these regions will have a resolution of 50 50 m. That is
four times higher resolution than the JERS-1 mosaics used here, so ALOS mosaics will contain a
minimum of four pixels for the smallest lake size class. This will undoubtedly increase the quality of
the estimates for the smallest lake classes. However, it may still be necessary to calibrate the mosaics
by comparison with small selected images of higher resolution data. This may also be possible for some
of the existing JERS-1 mosaics by comparison with some of the raw higher resolution JERS-1
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
300
K.H. TELMER AND M.P.F. COSTA
50,000
Sub-Arctic low estimate
Sub-Arctic high estimate
Central Canada low estimate
No. lakes per 100.000 km2
40,000
Central Canada high estimate
Pantanal low estimate
30,000
Pantanal high estimate
20,000
10,000
0
0.01-0.1 (1-10)
0.1-1 (10-100)
1-10 (100-1000)
Lake size range in
km2
10-100 (100-10000)
>100 (>10000)
(lake size range in pixels)
100,000
No. lakes per 100.000 km2
10,000
1,000
Sub-Arctic low estimate
100
Sub-Arctic high estimate
Central Canada low estimate
10
Central Canada high estimate
Pantanal low estimate
Pantanal high estimate
1
0.01-0.1 (1-10)
0.1-1 (10-100)
1-10 (100-1000)
Lake size range in
km2
10-100 (100-10000)
>100 (>10000)
(lake size range in pixels)
Figure 7. Estimates of lake size distribution for the three study areas based on classification of JERS-1 SAR imagery. The distribution
is shown in log scale in the top panel and in log–log scale in the bottom panel. The lake size ranges are given in km2 and in pixels. Each
pixel is 100 100 m or 0.01 km2. The digital number threshold used to separate lakes from land is higher for the high estimate (see text).
It is difficult to identify the small lakes that are represented by single or low numbers of pixels because these pixels sample a mix of land
and water. The distribution in the smallest size range is therefore probably low.
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
301
SAR-BASED ESTIMATES OF THE SIZE DISTRIBUTION OF LAKES
200,000
Carbon accumulation in lakes
(t/ 100.000 km2per year)
Based on imagery only
Based on imagery +
extrapolation for smallest lakes
150,000
100,000
50,000
0
Sub-Arctic Sub-Arctic
low
high
estimate
estimate
Central
Canada
low
estimate
Central
Canada
high
estimate
Pantanal
low
estimate
Pantanal
high
estimate
Figure 8. Estimates of the amount of carbon accumulated in lakes for the three regions.
imagery. More details about the problems of spatial resolution are discussed in Costa and Telmer
(2007).
It is also a strong possibility that the multi-polarimetric capabilities of ALOS will aid in imaging smaller
lakes. Influences of different interferences could be more effectively eliminated by using the increased
amount of information available from the ALOS mosaics. For example, HV polarization is more sensitive
to vegetation than HH and the two could be combined to reduce shoreline or aquatic vegetation
interferences. Similarly, the effect of wind on water surface roughness could be filtered out by combining
VV polarization. Combinations of these polarities may also vastly improve the ability to image lakes in the
Pantanal by allowing bare soils and grasslands to be better separated from water–the unimodal distribution
of JERS-1 data shown in Figure 3 suggests an almost continuous gradient of backscattering from water to
savannah to forest making threshold determination difficult.
The slope of the equations given in Table 1 also indicate which size category of lakes occupies the greatest
area. Slopes less than 1 (minus one) indicate that the smallest size category occupies the greatest area. This
is the case for the Arctic study area of Canada and for the Pantanal (Pantanal estimates need to be treated
cautiously). The opposite is true for the Central Canadian region. In other words small lakes are the most
important in terms of total water area for the Pantanal and Arctic study areas. They must therefore also be
the most important in terms of carbon accumulation for these regions. This is likely still to be true for the
Central Canadian region as well, if indeed small lakes accumulate carbon at a rate six times greater than
large ones as suggested by Pajunen (2000) and Kortelainen (2004).
Neither Pajunen nor Kortelainen estimated carbon accumulation rates for lakes smaller than 0.1 km2,
and so accumulation rates for these lakes are not well known. Considering that carbon accumulation
increases towards smaller lakes, it is probably reasonable to assume that the smallest lakes at least
accumulate carbon at the same rate as lakes >0.1 km2. Doing this and including the smallest size class of
lakes in carbon inventories for these landscapes greatly increases the amount of carbon accumulation
accounted for by lakes per unit area of landscape–by as much as 30% (see Table 2). This highlights the fact
that small lakes are important for carbon dynamics of landscapes, and that a better understanding of
carbon accumulation in these smallest of lakes is required. A reliable estimate of their number for various
regions is an important first step–one that that ALOS should be able to help with.
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
Arctic low
estimate
Arctic high
estimate
Copyright # 2007 John Wiley & Sons, Ltd.
71 097
11 250
15 238
21 615
19 487
3508
27%
26 825
11 250
6625
4913
3419
585
19 668
22.5
133
983
6838
11 693
Central Canada
low estimate
11 250
18 688
36 135
34 834
5951
31 331
106 857
132 237
11 250
8125
8213
6111
992
5222
34 783
38 921
35%
39%
22.5
163
1643
12 223
19 835
104 436
33 885
118 486
Central Canada
high estimate
65 096
6900
18 260
29 861
10 076
14%
14 207
3000
4150
5239
1679
44 953
60
830
10 478
33 585
Pantanal low
estimate
12 363
24 915
36 843
11 387
41 968
85 508
116 089
5375
5663
6464
1898
6995
19 551
24 496
20%
24%
108
1133
12 928
37 958
139 895
52 125
154 062
Pantanal high
estimate
nn
n
Underestimated values directly determined from imagery for lake size category 0.1–1 km2. These are underestimated for reasons shown in Figure 7 and explained in the text.
Extrapolated values (in italics) for lake size category 0.1–1 km2. This is done because the imagery is not capable of resolving the smallest size category of lakes. See
Figure 7 and the text. The relationships used for extrapolation were generated from the four largest lake size categories and are as follows:
Arctic: [number of lakes per 100 000 km2] = 1.09 * [lake size] + 3.92; R2 ¼ 0:992:
Central Canada: [number of lakes per 100 000 km2] = 0.92 * [lake size] + 3.82; R2 ¼ 0:999:
Pantanal: [number of lakes per 100 000 km2] = 1.04 * [lake size] + 3.79; R2 ¼ 0:999:
Carbon accumulation in t yr1 per size category per 100 000 km2
>100 (>10 000)
3750
3750
10–100 (100–10 000)
14 663
17 538
1–10 (100–1000)
35 640
41 085
0.1–1 (10–100)
30 139
39 572
0.01–0.1 (1–10)n
8255
12 056
0.01–0.1 (1–10)nn
65 727
TOTAL
92 447
114 000
TOTALnn
167 672
Area covered by lakes per size category per 100 000 km2
>100 (>10 000)
3750
3750
10–100 (100–10 000)
6375
7625
1–10 (100–1000)
8100
9338
0.1–1 (10–100)
5288
6943
0.01–0.1 (1–10)n
1376
2009
0.01–0.1 (1–10)nn
10 954
TOTAL
24 999
29 844
TOTALnn
38 609
% of land covered by lakes
25%
30%
% of land covered by lakesnn
39%
Numbers of lakes per size category per 100 000 km2
>100 (>10 000)
7.5
7.5
10–100 (100–10 000)
128
153
1–10 (100–1000)
1620
1868
0.1–1 (10–100)
10 575
13 885
0.01–0.1 (1–10)n
27 518
40 185
0.01–0.1 (1-10)nn
219 090
TOTAL
39 848
56 098
TOTALnn
235 002
Size category
km2 (pixels)
Table 2. High and low estimates of lake size distribution and carbon accumulation for the three study areas
302
K.H. TELMER AND M.P.F. COSTA
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
SAR-BASED ESTIMATES OF THE SIZE DISTRIBUTION OF LAKES
303
The estimates of lake distribution for large lakes in the Pantanal (Figure 7) are not correct. There are
fewer 10 to 100 km2 lakes per unit area in the Pantanal than are shown in Figure 8. The overestimate was
caused by confusion of intermittent floodways (‘vazantes’) and lakes. The floodways are relatively large
areas of scrubby grasslands and bare sand in the dry season that, as previously discussed, have a JERS-1
signature that overlaps with lakes. They were therefore erroneously classified as lakes using our
methodology. Similar results were reported with SEASAT L-band detection of lakes (Lewis, 1998). Again it
is likely that the multi-polarization and higher resolution data produced by ALOS-PALSAR will allow
these floodways to be differentiated from lakes more easily. Canada’s soon to be launched RADARSAT 2
satellite, which will acquire band C multi-polarization SAR, may also help with this task.
Similarly, the estimates for small lakes in the Pantanal appear to be too low according to the power law
trends. That may be because many lakes in the Pantanal contain tall-dense emergent aquatic vegetation
that causes radar images to appear brighter than they otherwise would because of the double-bounce
mechanism (Novo et al., 2002). Using our methodology, bright areas were classified as land and so any
lakes that expressed ‘double-bounce brightness’ were not included in the estimate.
There are no estimates of the rate of carbon accumulation for lakes in the Pantanal. This is a major gap
in understanding the carbon dynamics in the Pantanal and other similar environments. This information is
needed as it is known that these environments are highly productive and therefore carbon fluxes could be
large (Por, 1995; Pott and Pott, 2000). The estimates shown in Figure 8 for the Pantanal were produced
using the rates estimated by Kortelainen (2004) for Finnish lakes and are shown simply to illustrate that
lakes in the Pantanal may represent a significant carbon sink; however, their high productivity makes it
possible that accumulation rates are significantly higher than those shown. Improving these estimates will
require better imagery from ALOS and field-based studies to quantify the rate of carbon accumulation in
the Pantanal.
The carbon accumulation rates shown in Figure 8 for Central and Arctic Canada are based also on the
Finnish data and so the differences between the areas are purely due to the differences in the number and
size distribution of lakes. This shows that lake size and size distribution alone can potentially result in 25%
differences in carbon accumulation. Although these areas share much in common with areas in Finland,
field-based estimates of carbon accumulation as well as improved estimates of lake size distribution will be
required to improve confidence in these preliminary numbers.
REFERENCES
Abdon MM, Pott VJ, Silva JSV. 1998. Evaluation of aquatic vegetation on ponds of Nhecolândia subregion in the
Pantanal through Landsat and SPOT data. Pesquisa Agropecuaria Brasileira 33: 1675–1681.
Algesten G, Sobek S, Bergstrom AK, Agren A, Tranvik LJ, Jansson M. 2004. Role of lakes for organic carbon cycling
in the boreal zone. Global Change Biology 10: 141–147.
Chapman B, Siqueira P, Freeman A. 2002. The JERS Amazon multi-season mapping study (JAMMS): observation
strategies and data characteristics. International Journal of Remote Sensing 23: 1427–1446.
Costa M. 2004. Use of SAR satellites for mapping zonation of vegetation communities in the Amazon floodplain.
International Journal of Remote Sensing 25: 1817–1835.
Costa M. 2005. Estimate of net primary productivity of aquatic vegetation of the Amazon floodplain using Radarsat
and JERS-1 imagery. International Journal of Remote Sensing 26: 4527–4536.
Costa MPF, Telmer K. 2006. Utilizing SAR imagery and aquatic vegetation to map fresh and brackish lakes in the
Brazilian Pantanal wetland. Remote Sensing of the Environment 105: 204–213.
Costa M, Telmer K. 2007. Mapping and monitoring lakes in the Brazilian Pantanal wetland using synthetic aperture
radar imagery. Aquatic Conservation: Marine and Freshwater Ecosystems 17: 279–290.
Costa M, Niemann O, Novo E, Ahern F, Mantovani J. 2002. Biophysical properties and mapping of aquatic vegetation
during the hydrological cycle of the Amazon floodplain using JERS-1 and RADARSAT. International Journal of
Remote Sensing 23: 1401–1260.
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
304
K.H. TELMER AND M.P.F. COSTA
Dean WE, Gorham E. 1998. Magnitude and significance of carbon burial in lakes, reservoirs, and peatlands. Geology
26: 535–538.
Dobson MC, Pierce LE, Ulaby FT. 1996. Knowledge-based land-cover classification using ERS-1/JERS-1 SAR
composites. IEEE Transactions on Geoscience and Remote Sensing 34: 83–99.
Einsele G, Yan J, Hinderer M. 2001. Atmospheric carbon burial in modern lake basins and its significance for the
global carbon budget. Global and Planetary Change 30: 167–195.
Fee EJ, Hecky RE. 1992. Introduction to the Northwest Ontario Lake Size Series (NOLSS). Canadian Journal of
Fisheries and Aquatic Sciences 49: 2434–2444.
Fee EJ, Shearer JA, Debruyn ER, Schindler EU. 1992. Effects of lake size on phytoplankton photosynthesis. Canadian
Journal of Fisheries and Aquatic Sciences 49: 2445–2459.
Hall FG. 1999. Introduction to special section: BOREAS in 1999: experiment and science overview. Journal of
Geophysical Research, BOREAS Special Issue II 104: 27 627–27 640.
Hess LL, Melack JM, Filoso S, Wang Y. 1995. Delineation of inundated area and vegetation along the Amazon floodplain
with the SIR-C synthetic-aperture radar. IEEE Transactions on Geoscience and Remote Sensing 33: 896–904.
Jonsson A, Karlsson J, Jansson M. 2003. Sources of carbon dioxide supersaturation in clearwater and humic lakes in
northern Sweden. Ecosystems 6: 224–235.
Kelly CA, Fee E, Ramlal PS, Rudd JWM, Hesslein RH, Anema C, Schindler EU. 2001. Natural variability of carbon
dioxide and net epilimnetic production in the surface waters of boreal lakes of different sizes. Limnology and
Oceanography 46: 1054–1064.
Kortelainen P, Pajunen H, Rantakari M, Saarnisto M. 2004. A large carbon pool and small sink in boreal Holocene
lake sediments. Global Change Biology 10: 1648–1653.
Kozlenko N, Jeffries MO. 2000. Bathymetric mapping of shallow water in thaw lakes on the north slope of Alaska with
spaceborn imaging radar. Arctic 53: 306–316.
Leckie DG, Ranson KJ. 1998. Forestry applications using imaging radar. In Principles and Applications of Imaging
Radar, Manual of Remote Sensing, Vol. 2, 3rd edn, Henderson FM, Lewis AJ (eds). John Wiley: Chichester; 435–509.
Lewis AJ. 1998. Geomorphic and hydrologic applications of active microwave remote sensing. In Principles and
Applications of Imaging Radar, Manual of Remote Sensing, Vol. 2, 3rd edn, Henderson FM, Lewis AJ (eds). John
Wiley: Chichester; 567–629.
Martin S. 2004. Introduction to Ocean Remote Sensing. Cambridge University Press: Chichester.
Meybeck M. 1995. Global distribution of lakes. In Physics and Chemistry of Lakes, 2nd edn, Lerman A, Imboden D,
Gat J (eds). Springer Verlag: New York.
Mulholland PJ, Elwood JW. 1982. The role of lake and reservoir sediments as sinks in the perturbed global carbon
cycle. Tellus 34: 490–499.
Novo E, Costa M, Mantovani J, Lima I. 2002. Relationship between macrophyte stand variables and radar backscatter
at L and C band–Tucuruı́ reservoir–Brazil. International Journal of Remote Sensing 23: 1241–1260.
Pace ML, Cole JJ, Carpenter SR, Kitchell JF, Hodgson JR, Van de Bogert MC, Bade DL, Kritzberg ES, Bastviken D.
2004. Whole-lake carbon-13 additions reveal terrestrial support of aquatic food webs. Nature 427: 240–243.
Pajunen H. 2000. Lake sediments; their carbon store and related accumulation rates. In Carbon in Finnish Lake
Sediments, Pajunen H (ed.). Special Paper no. 29, Geological Survey of Finland, 39–69.
Por FD. 1995. The Pantanal of Mato Grosso (Brazil). In World’s Largest Wetlands, Dumont HJ, Werger MJA (eds).
Kluwer: Drodrecht.
Pott VJ, Pott A. 2000. Plantas Aquaticas do Pantanal, EMBRAPA.
Rosenqvist A, Shimada M, Chapman B, Freeman A, De Grandi G, Saatchi S, Rauste Y. 2000. The global rain forest
mapping project–a review. International Journal of Remote Sensing 21: 1375–1387.
Schindler DW, Curtis PJ, Parker BR, Stainton MP. 1996. Consequences of climate warming and lake acidification for
UV-B penetration in North American boreal lakes. Nature 379: 705–708.
Schlünz R, Schneider RR. 2000. Transport of terrestrial organic carbon to the oceans by rivers: re-estimating flux- and
burial rates. International Journal of Earth Science 88: 599–606.
Simard M, De Grandi G, Saatchi S, Mayaux P. 2002. Mapping tropical coastal vegetation using JERS-1 and ERS-1
radar data with a decision tree classifier. International Journal of Remote Sensing 23: 1461–1474.
Smith LC, Sheng Y, MacDonald GM, Hinzman LD. 2005. Disappearing arctic lakes. Science 308: 1429.
Ulaby FT, Moore RK, Fung AK. 1982. Microwave Remote Sensing: Radar Remote Sensing and Surface Scattering and
Emission Theory, Vol. II. Artech House: Norwood, MA.
Ulaby FT, Moore RK, Fung AK. 1986. Microwave Remote Sensing, from Theory to Applications, Vol. 3. Artech House:
Norwood, MA.
Copyright # 2007 John Wiley & Sons, Ltd.
Aquatic Conserv: Mar. Freshw. Ecosyst. 17: 289–304 (2007)
DOI: 10.1002/aqc
Документ
Категория
Без категории
Просмотров
6
Размер файла
987 Кб
Теги
base, investigation, distributions, lake, sar, estimates, brazil, tool, size, canada, carbon
1/--страниц
Пожаловаться на содержимое документа